Grok in vibe coding: a test with a 3D scene, a game, and website design

The author tested Grok 4.7 on four tasks: creating a 3D orc model, an aquarium, a Super Mario game, and a website design. He evaluated the appearance and operation of interactive elements, while also recording the time and cost of each run.

The results belong to the author of the test.

Conditions of the Grok test

The four projects took a total of 97 minutes 21 second, and the stated cost was $11,43. This sum consists of the expenses for the four Grok tasks; the cost of a separate comparative run with another model was not specified. The author provided the duration and expenses for each project, but did not describe the cost-calculation method, environment settings, or exact prompts. The figures reflect his runs, not a standard generation price.

These tasks test the generation of code and visual prototypes, but do not replace a systematic comparison of models. The result depends on the wording of the prompt, scene requirements, execution environment, and time spent on fixes. The author did not disclose these conditions in detail, so the available data cannot separate the model’s influence from the task setup and subsequent refinement.

Grok and the 3D orc model

The orc model took 22 minutes, with a stated cost of $2,87. The author considered the result unfit for further use: in his assessment, the character’s teeth were missing, and he held the weapon incorrectly. The strike looked roughly right, but this successful element did not make up for the errors in the details and pose.

In this run, the author evaluated not only the presence of an object on the screen, but also the readability of the silhouette, the position of the hands and weapon, and whether the details matched the description. He noted a gap between an individual movement functioning properly and the model’s overall visual plausibility. This is an assessment of a specific scene, not a conclusion about the quality of all Grok 3D models.

One orc is not enough to conclude that Grok always builds 3D models poorly. In this run, the author assessed the scene as unfit for use.

Grok aquarium and comparative run

The aquarium took 30 minutes to generate and cost the stated $3,35. The author called the result unsuccessful: the rocks looked strange, the crab moved sideways, its claws remained still, and the fish had no noticeable tails or fins. The author did not refine the project further, citing the expense.

In this scene, the author evaluated individual objects and their movement. He noted that the fish lacked characteristic details and that the crab’s animation looked incorrect. In his assessment, the presence of objects in the frame did not make up for these shortcomings.

In a separate comparative run, the author used the GPT-6.1 Sol model in High mode and rated the aquarium positively. He listed fish, water reflections, lighting, particles, bubbles, moving crabs and snails, a castle, and plants. According to him, no additional 3D was used.

This comparison reflects a personal assessment of two results, not a controlled experiment. The description contains no data to verify whether the prompts, environment, settings, and number of attempts were the same. The conclusion is limited to the author’s observation: his GPT-6.1 Sol run seemed more successful than the specific Grok aquarium. The comparison does not establish which model handles aquariums better overall.

Super Mario game: Grok made the project playable

The game took 27 minutes and cost $3,08.

The author reported that it was playable, meaning the result worked as a game prototype. His main complaint concerned the character’s appearance: in the author’s assessment, Mario looked like a cyclops and more closely resembled a leprechaun.

This assessment contains two separate criteria. The project performed the basic game function, but the character’s appearance did not meet the author’s expectations. A result can work as a prototype while still requiring the graphics and hero character design to be redone.

The author did not provide a detailed check of the controls or set of game features. Therefore, the test conveys his overall impression of playability but does not confirm the project’s completeness.

Grok website design: interactivity and readability

The website design took 18 minutes 21 second and cost $2,13. The author found the interface difficult to understand: he considered some elements and labels hard to read and described the composition as abstract. At the same time, some elements were clickable, and the sliders changed the 3D object.

In this example, functionality and interface usability received different assessments. The author was able to click some elements and change the object with the sliders, but the labels and composition seemed unclear to him. When assessing the page, he considered interactivity, readability, and ease of understanding the controls separately.

The author did not say whether he conducted user testing or evaluated the layout himself. His comments describe an impression of a specific page, not the results of a usability study.

What the Grok vibe-coding test shows

After the four tasks, the author called Grok 4.7 expensive and weak for vibe coding and saw no reason for himself to continue using the model through a subscription or API. This is a personal conclusion based on a small set of runs, not a universal ranking or assessment of all programming scenarios.

Before paying, test the model on a task similar to yours. Decide in advance what will count as a finished result: functioning mechanics, an accurate 3D form, a readable interface, or a combination of these requirements. Compare generation time and the amount of manual fixes. Even the quick prototype in this test would still have required evaluation and possibly further refinement.

The test does not provide the original prompts, detailed evaluation criteria, or repeat attempts for each task. Without this data, it is impossible to separate the model’s influence from the task setup and refinement process. The cost of one run cannot be applied to future work through an API or subscription: such a calculation requires the pricing terms and a method for accounting for expenses. The author also did not specify how much time would have been needed to fix each project.

The author suggested that another model could create ten aquariums for the price of one Grok project and do a better job. This is an unverified assumption, not a measurement result.

FAQ about Grok in vibe coding

Can this test be used to call Grok useless?

No. The test describes four specific runs and the author’s personal criteria. The game turned out to be playable, and some controls worked on the website. These observations do not prove that the model is unsuitable for other tasks or prompts.

Which Grok result turned out to work best?

The author called the Super Mario game playable. Some elements on the website were clickable, and the sliders changed the 3D object. However, he had comments on both results: the character’s appearance did not meet his expectations, and the website seemed difficult to read.

Should you choose Grok based on this comparison with GPT-6.1 Sol?

Use the comparison only as a reason to conduct your own test. It concerns an aquarium and is based on the author’s assessment; there is not enough data for a rigorous comparison of the conditions. When deciding on a subscription or API, repeat the test on your own task and account for the time, cost, and subsequent manual work.